# AI Agent Risk Checklist

Use this before an agent touches customers, money, private data, or important records.

## Risk Categories

Score each from 1 low to 5 high.

| Risk | Score | Notes |
| --- | ---: | --- |
| Customer impact |  |  |
| Financial impact |  |  |
| Legal/compliance impact |  |  |
| Privacy/data exposure |  |  |
| Brand/reputation impact |  |  |
| Operational reversibility |  |  |

## Approval Rules

Require human approval when the agent:

- sends customer messages
- changes CRM stages
- updates billing
- issues refunds
- edits legal/compliance language
- handles private data
- gives high-stakes recommendations
- is uncertain or missing source context

## Launch Modes

Start here:

1. recommendation only
2. draft for review
3. approved action
4. sampled review
5. low-risk automatic action

Most first workflows should start at 1 or 2.

## Monitoring

Track:

- reviewer edits
- rejected outputs
- failed tool calls
- escalations
- customer complaints
- wrong source usage
- time saved after review

Related A8gent pages:

- `/tools/ai-agent-risk-scorer`
- `/tools/ai-agent-readiness-quiz`
- `/courses/ai-agent-implementation-course`